SamanthaStorm/tether-darvo-regressor-v1

Tether Pro DARVO Regressor v2

Model Description

This model detects DARVO (Deny, Attack, Reverse Victim & Offender) manipulation tactics in text communication. DARVO is a psychological manipulation strategy where an abuser:

1. Denies the abuse ever happened
2. Attacks the victim for bringing it up
3. Reverses the roles to claim they are the victim

Key Features

🎯 Role-Aware Detection: Distinguishes between genuine accountability and manipulation tactics 🔬 Research-Grade Accuracy: 84% accuracy with 0.88 AUC ⚡ Real-Time Analysis: Optimized for fast inference 🛡️ Professional Use: Designed for therapists, legal professionals, and safety applications

Performance Metrics

Metric	Score
R²	0.665
MAE	0.171
MSE	0.043
Accuracy	84.2%
AUC	88.1%

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("SamanthaStorm/tether-darvo-regressor-v1")
model = AutoModelForSequenceClassification.from_pretrained("SamanthaStorm/tether-darvo-regressor-v1")

# Example usage
text = "You're the one being abusive to me right now"

# Tokenize and predict
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)
with torch.no_grad():
    outputs = model(**inputs)
    darvo_score = outputs.logits.item()

print(f"DARVO Score: {darvo_score:.3f}")  # Higher scores = more DARVO tactics

Score Interpretation

• 0.0 - 0.3: Genuine accountability, healthy communication
• 0.3 - 0.6: Some defensive patterns, mild deflection
• 0.6 - 0.8: Moderate DARVO tactics, concerning patterns
• 0.8 - 1.0: Strong DARVO tactics, victim reversal

Example Predictions

Text	DARVO Score	Interpretation
"You're the one being abusive to me right now"	0.870	High DARVO - victim reversal
"I don't remember saying that"	0.224	Low DARVO - simple denial
"I take full responsibility for my actions"	0.205	Very low DARVO - accountability

Training Data

Trained on 285 carefully curated examples including:

• High DARVO: Explicit victim reversal tactics
• Medium DARVO: Deflection and minimization patterns
• Low DARVO: Genuine accountability and healthy communication
• Contrast Examples: Non-apologies vs real apologies

Applications

🏥 Clinical Therapy

• Help therapists identify manipulation patterns in client relationships
• Assist in couples counseling to recognize unhealthy dynamics
• Support trauma therapy by validating victim experiences

⚖️ Legal Documentation

• Analyze communication patterns in domestic violence cases
• Provide objective evidence of psychological manipulation
• Support legal professionals in building abuse cases

🏢 Workplace Safety

• Identify harassment patterns in workplace communications
• Support HR investigations with objective analysis
• Create safer work environments through pattern recognition

Ethical Considerations

⚠️ Important: This model is designed to assist professionals and should not be used as the sole basis for serious decisions about relationships or safety.

• Professional Use: Best used by trained therapists, counselors, and legal professionals
• Context Matters: Consider cultural, situational, and individual factors
• Not Diagnostic: Does not diagnose psychological conditions
• Privacy: Ensure consent when analyzing personal communications

Technical Details

• Base Model: DistilBERT (distilbert-base-uncased)
• Architecture: Custom regression head with 4-layer neural network
• Training: 8 epochs with cosine learning rate scheduling
• Optimization: Mixed precision training (FP16)
• Max Length: 256 tokens for efficiency

Model Architecture

DistilBERT Base
    ↓
Linear(768 → 768) + GELU + Dropout
    ↓  
Linear(768 → 384) + GELU + Dropout
    ↓
Linear(384 → 192) + GELU + Dropout
    ↓
Linear(192 → 1) + Sigmoid
    ↓
DARVO Score (0.0 - 1.0)

Version History

v2 (Current)

• ✅ Enhanced training dataset (285 examples)
• ✅ Improved architecture with deeper regression head
• ✅ Better score calibration for accountability detection
• ✅ Added contrast examples (fake vs real apologies)
• ✅ 84% accuracy (up from 40%)

v1 (Previous)

• Basic DARVO detection capability
• Limited training data
• Lower accuracy performance

Citation

If you use this model in research or professional practice, please cite:

@misc{tether-darvo-regressor-v1,
  title={Tether Pro DARVO Regressor: Role-Aware Detection of Manipulation Tactics},
  author={SamanthaStorm},
  year={2024},
  howpublished={\url{https://huggingface.co/SamanthaStorm/tether-darvo-regressor-v1}},
}

Contact & Support

For questions about integration, licensing, or professional applications:

• 📧 Enterprise: [email protected]
• 🌐 Documentation: docs.tether.ai
• 📅 Consultation: calendly.com/tether-pro

Related Models

Part of the Tether Pro AI Suite:

• 🛡️ Boundary Health Detector: SamanthaStorm/healthy-boundary-predictor
• 🎯 Abuse Pattern Detector: SamanthaStorm/tether-multilabel-v6
• 🎭 Sentiment Analyzer: SamanthaStorm/tether-sentiment-v3
• 🧩 Fallacy Detector: SamanthaStorm/fallacy-detector (coming soon)
• 🎯 Intent Classifier: SamanthaStorm/intent-detector (coming soon)

---

Built with ❤️ for safer communication analysis